Prakamya Mishra

I'm an applied research engineer at AMD AI in Seattle, where I am part of the GenAI team lead by Zicheng Liu, working on novel techniques to efficiently train LLMs & image/video generation AI models on large-scale clusters.

Previously, I was a graduate researcher in BioNLP Lab at UMass Amherst (Mentors: Hong Yu & Zonghai Yao), graduate research with Amazon Science NLU Team (Mentor: Mayank Kulkarni), and an applied scientist intern at Amazon Science SLU Team (Mentors: Abhishek Malali & Masha Belyi). I did my Masters in Computer Science from UMass Amherst.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  Twitter  /  Github

Research

I'm interested in natural language processing, deep learning, and generative AI. Most of my research is about language models, representation learning and applications of NLP.

Introducing TTT-Bench, a new benchmark that is designed to evaluate basic reasoning abilities in LRMs through a suite of four two-player Tic-Tac-Toe-style games.

Announcing Instella-Long, a long-context language model continually trained from Instella-3B-Instruct on AMD Instinct™ MI300X GPUs.

Announcing Instella, a series of 3 billion parameter language models developed by AMD, trained from scratch on 128 Instinct MI300X GPUs.

AMD OLMo are a series of 1 billion parameter language models pre-trained with 1.3 trillion tokens on 16 nodes, each with four (4) AMD Instinct™ MI250 GPUs.

This study leverages synthetic edit feedback to improve factual accuracy in clinical summarization using DPO and SALT techniques. Our approach demonstrates the effectiveness of GPT-generated edits in enhancing the reliability of clinical NLP applications.

We propose a novel clustering-based few-shot sampling approach that leverages intrinsically available sub-domain information as topics from the dataset to extract few-shot samples to be labeled from the target domains and be used for fine-tuning.

In this work, we propose a new pipeline using ChatGPT instead of human experts to generate high-quality feedback data for improving factual consistency in the clinical note summarization task.

In this work, we present a novel multi-modal deep neural network architecture that uses speech and text entanglement for learning phonetically sound spoken-word representations.

We propose NeuralNERE, an end-to-end Neural Named Entity Relationship Extraction model for constructing climate change knowledge graphs directly from the raw text of relevant news articles. Additionally, we introduce a new climate change news dataset (called SciDCC dataset) containing over 11k news articles scraped from the Science Daily website.

Developed novel Bi-directional Inter-Sentence Contextual Attention mechanism (Bi-ISCA) to capture inter-sentence dependencies for detecting sarcasm. Explained model behaviors and predictions by analyzing the attention maps and identifying words responsible for invoking sarcasm.

Miscellanea

Design and source code from Jon Barron's website.